Posted by B. Douglas Ward on March 15, 2001 at 16:02:01:
In Reply to: Re: 3dDeconvolve for t-test posted by Phoebe on March 15, 2001 at 14:30:08:
Phoebe:
1) The -rmsmin option was intended primarily to speed up program execution
by excluding voxels with little temporal variation. The more recently added
-mask option is better. Such a mask does not have to be very precise, as
long as it does not exclude voxels that are inside the brain. You could
produce such a mask by using a simple intensity threshold (say, using the
3dcalc step function). If processing time is not an issue, then you can
dispense with both the -rmsmin and -mask options.
Program 3dIntracranial is for generating high resolution (anatomical) masks,
and is not appropriate for generating low resolution functional masks.
Since the 3dDeconvolve calculations are performed independently for each
voxel, masking voxels outside the brain has absolutely no effect on the
results for voxels inside the brain. I say this just to avoid any possible
confusion.
2) You're correct. The Func subbrick determines the color coding in the
display, and the Thr subbrick determines which voxels "light up". For your
experiment, the logical choice is to use " 'para t^1 coef' for function
sub-brick and 'para t^1 t-st' for threshold sub-brick."
3) Regarding the experimental design itself: I don't know the details and
considerations behind the design that you are using. However, let me suggest
that you consider some alternative to the periodic block design that you
described. Periodic block designs have a number of limitations. The long
"on" periods can result in habituation or lack of attention. The periodicity
of the design can result in accidental confounding of the response with natural
biological processes, not to mention confounding with subject motion. The fixed
nature of the design can also lead to anticipatory effects.
Much of the information that comes from the fMRI time series data is due to
the changes that occur in transition from one state to another state. But for
your design, which has very long blocks, there are very few transitions. If
we consider the response in one block to be a single measurement, then your
entire time series effectively contains only 2 measurements, not 120.
The point of all this is: I recommend that you randomize the stimulus
presentation. If the stimulus must occur in blocks, then try to use much
shorter blocks, and randomize the placement of the blocks within the given
time constraints. Please read the documentation contained in file
3dDeconvolve.ps. Pay particular attention to the examples in Section 1.4.1:
Evaluation of the Experimental Design, and Section 3: Program RSFgen.
With a randomized stimulus presentation, you will probably want to use
the -stim_maxlag option to model the hemodynamic response. This is also
described in the above documentation.
Doug Ward